Trend-following - Deep learning can help but burden still on setting up the problem correctly

Deep learning techniques are a natural for quantitative investing and more specifically for trend-following, but just because the concept of deep learning is good for finding improved models does not mean it should be used. I am not anti-deep learning but realize that there are high barriers to entry to get it right as well as barriers to success based on costs of transacting. There is no simple answer with using these techniques because the value-added is all conditional on the set-up of the problem and the type of techniques used. 

A recent paper applying deep learning to trend-following shows that it can have a significant improvement in Sharpe ratio and other key statistics, yet there can be high variation in success based on problem formatting. See "Enhancing Time Series Momentum Strategies Using Deep Neural Networks". 

No one should be surprised with this result. There have been hundreds of trend-followers through time that have all had their own variation on trend identification, position sizing, and portfolio construction. There is no single model solution. There is high variation in performance based on the techniques used and assumptions made. 

The researchers form a control strategy without any deep learning and then compare against four different approaches: 

Lasso regression (LINEAR)

Multi-layer perceptron (MLP)

Wavenet based on convolutional networks (CNN) 

Long short-term Memory (LSTM).

The techniques employ increasing complexity with LINEAR being a relatively simple linear model, MLP using a 2-layer neural net, wavenet  accounting for long history, and LSTM using sequence prediction based on memory structures.

These are some of the best techniques for use with time series data and are fitted against a number of objective functions such as Sharpe ratio, return, MSE, and binary choice. All returns are rescaled to a target volatility to ensure fair comparisons. The following two exhibits show the cumulative returns for the set of models. The box plots show the range of results across individual assets. There can be large differences between asset return performance across models.

Deep learning techniques can improve gross performance statistics, but it comes at a cost. More turnover translates to greater drag based on the transaction costs of constant adjustment. The impact of transaction costs is not trivial and can turn the great strategy into a marginal or losing model. The exhibits below show the added turnover and the impact on Sharpe ratio as transaction costs increase. 

Deep learning or any strategy with constant adjustments work effectively when transaction costs are not included or are very low. Increases in cost will have a much greater impact on deep learning models than simple approaches. Hence, transaction costs must be accounted for when building models. 

The challenges of trend-following research are the same across time. Improvements can be made through better and new techniques but costs matter and may lead the investor back to simple models.